CN103471709A - Method for predicting noise quality of noise inside passenger vehicle - Google Patents

Abstract

The invention discloses a method for predicting noise quality of noise inside a passenger vehicle. The problem that subjective evaluation of sound quality in the vehicle consumes time and is complicated and the work amount is large is solved. The method comprises the steps of 1 utilizing artificial heads or two microphones to acquire an interior noise signal; 2 adopting an auditory peripheral calculation model to calculate A-noise-level and noise quality objective psychological acoustic parameters; 3 performing subjective evaluation of interior noise quality, namely performing subjective evaluation of noise quality of a noise sample to obtain a subjective evaluation result of tested data; 4 establishing an objective quantitative model of the interior noise quality subjective evaluation through a BP neural network improved by means of a genetic algorithm based on the interior A-noise-level and noise quality objective psychological acoustic parameters and the noise quality subjective evaluation result; 5 re-establishing an objective quantitative model of the interior noise quality subjective evaluation based on objective parameters having important influence and the noise quality subjective evaluation result; 6 establishing an objective psychological acoustic parameter database of the interior noise quality.

Description

Passenger car internal car noise sound quality prediction method

Technical field

The present invention relates to a kind of internal car noise sound quality prediction method.In particular, the present invention relates to a kind of passenger car internal car noise sound quality prediction method.

Background technology

Along with the continuous progress of science and technology, people not only require automobile " peace and quiet ", also are concerned about the frequency constituent of internal car noise, and the factors such as the comfortableness of sound consider the quality of sound.In the orthodox car noise control technique, only emphasize the size of noise magnitude, think that noise level is more low better.But because people's subjective factor is got involved, there will be that to measure with dB (A) be up to standard, but feel to be harassing and wrecking property, with dB (A), measure large sound, and sensuously little more melodious than sound.Therefore, the vehicle completed under conventional art instructs, environment inside car is often too dull to be constrained, and subjective feeling is poor, has a strong impact on the comfortableness in car, restricts its strive power competing in market.

In-vehicle sound quality is under the requirement that meets human and environment, estimates the subjective feeling of the people of main body to internal car noise, and it can reflect the overall experience of passenger to sound in car comprehensively.Evaluation to in-vehicle sound quality is divided into subjective assessment and objective evaluation.Subjective estimate method adopts the forms such as survey or subjective assessment test, to the subjective assessment data by methods such as statistical study, obtain suitable evaluation term to describe the subjective perception feature of in-vehicle sound quality, test findings and people's subjective feeling has good consistance, and shortcoming is that whole process is loaded down with trivial details and workload is larger.Method for objectively evaluating is to obtain the parameter of Vehicle Interior Noise by the method for analyzing and measure, and estimates the quality of internal car noise, and objectivity is stronger, but has ignored people's subjective feeling.Now, associated mechanisms all is devoted to set up the subjective assessment of in-vehicle sound quality and the relation between the objective examination, not only qualitative but also describe quantitatively in-vehicle sound quality.But above method has limitation, can't system in-vehicle sound quality is carried out to prediction and evaluation.Accurately and efficiently carry out the in-vehicle sound quality prediction, all significant for evaluation, analysis and the control of car acpistocs design optimization, internal car noise.Therefore, be badly in need of at present a kind of internal car noise sound quality evaluation Forecasting Methodology, the design that is in-vehicle sound quality in the prototype design stage of automobile product provides foundation.

Summary of the invention

Technical matters to be solved by this invention is to have overcome prior art to have in-vehicle sound quality subjective assessment loaded down with trivial details, problem that workload is large consuming time, and a kind of passenger car internal car noise sound quality prediction method is provided.

For solving the problems of the technologies described above, the present invention adopts following technical scheme to realize:

The step of described passenger car internal car noise sound quality prediction method is as follows:

1) utilize dummy head or two microphones to gather the internal car noise signal:

Dummy head or two microphones are placed in to the copilot position, the internal car noise signal of collection vehicle under operating mode at the uniform velocity, and the internal car noise signal under operating mode is at the uniform velocity carried out to pre-service;

2) calculate the objective psychoacoustic parameter of A sound level harmony quality:

Adopt computational modeling of auditory periphery to calculate the objective psychoacoustic parameter of in-vehicle sound quality;

3) in-vehicle sound quality subjective assessment:

Noise sample is carried out to the subjective assessment of sound quality, obtain the subjective assessment result of test data;

4) set up the objective quantification model of in-vehicle sound quality subjective assessment:

Take A sound level in car, the objective psychoacoustic parameter of in-vehicle sound quality and sound quality subjective assessment result as basis, by the improved BP neural network of genetic algorithm, set up the objective quantification model of the subjective assessment of in-vehicle sound quality;

5) weight analysis and reconstruction evaluation model:

Find out the objective parameter that the subjective assessment result is had to material impact by weight analysis, take the objective parameter of material impact and sound quality subjective assessment result is basis, rebuilds the objective quantification model of the subjective assessment of the sound quality under steady state condition;

6) set up the objective psychoacoustic parameter database of in-vehicle sound quality:

In concrete test, set up the database of the objective psychoacoustic parameter of sound quality of different automobile types, as the data basis of whole passenger car internal car noise sound quality prediction method; Database is convenient to solve in-vehicle sound quality prediction mark is analyzed in the vehicle acoustic design is optimized, and further improves sound quality prediction appraisement system, for passenger car internal car noise sound quality prediction method is set up the basic platform of a system.

The step of utilizing dummy head or two microphone collection internal car noise signals described in technical scheme is as follows:

1) dummy head or two microphones are placed on to the copilot place, the vertical coordinate that is the normal sitting position of dummy head's position behaviour, two microphone particular locations is that seat surface center line and horizontal ordinate are the above 0.7 ± 0.05m of backrest surface center line intersection point places, and two microphones are arranged in 0.2 ± 0.02m place, plane of symmetry left and right of seat surface and backrest symmetrically;

2) tested object is selected the basis of the passenger car of different model different brands as the collecting test sample, quantity with the passenger car model is as the criterion, passenger car brand and model contain main flow passenger car vehicle on market, guarantee the sample car of 3～4 kinds of brands of model selection of the same race, as A, B and the C level car of car;

The working condition requirement while 3) testing, noise signal gathered, transmission for vehicles is put the most high-grade, take 10km/h～20km/h as a step-length, from the idling to the top speed, at the uniform velocity travel respectively, record the noise level of each speed of a motor vehicle in getting off, each working condition requirement record 2 times, be greater than 30s writing time;

Sample sound to the internal car noise that gathers under each steady state condition is processed, and needs through segmentation, intercepting and screening, and the length legislations of each operating mode sample sound intercepting is the 5s left and right, and requires to be subject to the steady state noise signal of external interference minimum as far as possible;

More its last result of calculation of test sample book quantity is more accurate, considers the requirement of testing cost and method computation complexity, and quantity, in 70～150 left and right, at least guarantees 30.

The step of the objective psychoacoustic parameter of the employing computational modeling of auditory periphery calculating in-vehicle sound quality described in technical scheme is as follows:

(1) extract two microphones and gather the left and right otoacoustic emission frequency spectrum of arbitrarily downgrading, and it is carried out to frequency range analysis, calculate binaural sound by formula (2) and arbitrarily downgrade:

In formula, L _{p, L/R}(f) be binaural effect sound pressure level dB; The summation of A (p, f) left and right ear loudness; The absolute acoustic pressure P of p; F signal frequency Hz; L _p,Lthe left otoacoustic emission dB that arbitrarily downgrades; L _p,Rauris dextra sound pressure level dB.

Determine the vehicle interior sound field of each frequency band, and determine whether reverberation field, if reverberation field, carry out the reverberation field attenuation processing, carry out again main loudness calculating, if not, with regard to direct main loudness, calculate, slope loudness is calculated and is finally derived ears characteristic loudness and total loudness, and its ears characteristic loudness is as follows:

$N^{\′}\left(z\right)=0.08\left(\frac{E_{\mathrm{T\Ω}}}{E_0}\right)[{(0.5+0.5\frac{E}{E_{\mathrm{T\Ω}}})}^{0.23}-1]---\left(3\right)$

In formula, E _{t Ω}for excitation corresponding to the threshold of audibility under quiet situation; E ₀for reference sound intensity; I ₀=10 ^-12w/m ²corresponding excitation; E is calculated the excitation that sound is corresponding;

Ears loudness computation model adopts the computation model of Zwicker, uses 1/3 octave band as basic data, and introduce the critical band concept masking effect of people's ear is done to corresponding being modified to,

$N=\underset{0}{\overset{24\mathrm{Bark}}{\∫}}{N}^{\′}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, be N'(z) characteristic loudness in critical band, unit is sone _g/ Bark, subscript G representation feature loudness value is calculated and is got by the critical band sound level;

(2) the calculating sharpness is as follows:

$S=k\×\frac{{\∫}_0^{24}{N}^{\′}\left(z\right)g\left(z\right)\mathrm{zdz}}{N}\mathrm{acum}---\left(5\right)$

In formula, S represents sharpness, the acum of unit; K=0.11 is weighting coefficient; N is ears loudness; N ' is (z) characteristic loudness in critical band; Z is critical band Bark number; G (z) is the weighting function of Zwicker according to the different critical frequency band, and its analytic expression about the Bark territory of g (z) is

$g\left(z\right)=\left\{\begin{array}{ccc}1& \mathrm{for}& z\≤16\\ {0.066e}^{0.171z}& \mathrm{for}& z>16\end{array}\right\}---\left(6\right)$

Z is critical band Bark number.

(3) the calculating roughness is as follows:

$R=0.3f_{\mathrm{mod}}{\∫}_{z=0}^{24\mathrm{Bzrk}}{\mathrm{\ΔL}}_E\left(z\right)\mathrm{dzasper}---\left(7\right)$

In formula: R represents roughness, the asper of unit; f _modfor modulating frequency; Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage, be defined as:

${\mathrm{\ΔL}}_E\left(z\right)={20\log }_{10}\left(\frac{{N^{\′}}_{\max }\left(z\right)}{{N^{\′}}_{\min }\left(z\right)}\right)---\left(8\right)$

In formula: N ' _maxand N ' (z) _min(z) maximal value and the minimum value of difference representation feature loudness.

(4) calculating shake degree is as follows:

$F=0.008\frac{{\∫}_0^{24}\mathrm{\ΔL}\left(z\right)\mathrm{dz}}{\left(\frac{f_{\mathrm{mod}}}{4}\right)+\left(\frac{4}{f_{\mathrm{mod}}}\right)}\mathrm{vacil}---\left(10\right)$

In formula: f _modfor modulating frequency, Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage;

(5) analysis that the tone degree passes through each frequency band, in conjunction with the left and right feature band, obtains the related coefficient of adjacent feature frequency band, and obtains after calculating modulating frequency, and its computation model is as follows

$T=\sqrt{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left[W_1\left({\mathrm{\Δz}}_i\right)W_2\left(f_1\right)W_3\left({\mathrm{\ΔL}}_i\right)\right]}^2}---\left(11\right)$

In formula, T is the tone degree, and unit is tu; W ₁(Δ Z _i) mean the relation of i single-frequency composition and critical band difference; W ₂(f _i) mean the relation of i single-frequency composition and frequency; W ₃(Δ L _i) mean the sound level surplus graded effect of i single-frequency composition;

(6) obtain the AI index according to the analysis to each frequency band, speech intelligibility is sound pressure level and the frequency (1/3rd octaves) that the AI index depends on ground unrest, the speech range that when people talk in car, people's ear can be listened shows as the zone of a 200～6300Hz in the spectrogram of ground unrest, and its computation model is

AI=∑W(f)D(f)/3

$\left\{\begin{array}{c}N\left(f\right)>\mathrm{UL}\left(f\right),D\left(f\right)=0\\ \mathrm{LL}\left(f\right)<N\left(f\right)<\mathrm{UL}\left(f\right),D\left(f\right)=\mathrm{UL}\left(f\right)-N\left(f\right)\\ N\left(f\right)<\mathrm{LL}\left(f\right),D\left(f\right)=30\end{array}\right.---\left(12\right)$

In formula, W (f) is the weighted coefficient; N (f) is noise spectrum; UL (f) is the ground unrest upper limit, and its expression formula is

lL (f) is the ground unrest lower limit, and its expression formula is

The step of the objective quantification model of setting up the in-vehicle sound quality subjective assessment described in technical scheme is as follows:

1) determine initial configuration and the parameter of BP NEURAL NETWORK;

In concrete test, recommend 3 layers of neural network to build the objective quantification model of in-vehicle sound quality subjective assessment, input layer is the input vector X that shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and 7 objective parameters of AI index form, and nodes is 7; Output layer is subjective sound quality evaluation score value, and nodes is 1; The hidden layer node number is mainly by the purposes decision of network, and suitable number does not have clear and definite theory provision, and the experimental formula of reference is as follows:

$m=\sqrt{n+l}+\mathrm{\&alpha;}---\left(13\right)$

m=log ₂n （14）

$\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{C}_m^i>k---\left(15\right)$

In above-mentioned 3 formulas: m is the hidden layer neuron number; N is the input layer number; L is the output layer neuron number; A, i are constant; K is sample number; Hidden layer adopts the logsig function, and output layer adopts the purlin linear function to build artificial neural network BP model;

2) determine BP NEURAL NETWORK operation, parameters, encoded

Introduce the Genetic Algorithms operative algorithm BP NEURAL NETWORK is optimized, select the encoding scheme of genetic algebra, variation probability, population scale and employing; The length of chromosome coding is by each layer of contained neuron number decision, and the input layer of establishing BP NEURAL NETWORK has R neuron, and hidden layer has S1 neuron, and output layer neuron number is S2, code length S=R*S1+S1*S2+S1+S2;

3) carry out the Genetic Algorithms operation according to fitness

Fitness function determines the learning error of population training by sample, and learning error is expressed as follows:

$E\left(i\right)=\underset{n}{\mathrm{\&Sigma;}}\underset{m}{\mathrm{\&Sigma;}}{(y_m-o_m)}^2---\left(16\right)$

In above formula, i is the chromosome number, and m is the output node number, and n is number of training.The fitness function of Genetic Algorithms can be expressed as:

f(i)=1/E(i) （17）

In above formula, f (i) is each individual fitness function;

Utilize the GA tool box of Matlab to be solved, and the individual optimal solution obtained is decoded, as the initial network weights of neural network, the corresponding relation of each real number and BP NEURAL NETWORK weights, threshold value during decoding;

4) using the optimization weights that obtain as the initial network weights, then use genetic algorithm neural network training BP;

After adopting genetic algorithm optimization acquisition initial weight and threshold value, just can carry out neural metwork training by setup parameter, adopt the additional momentum method to be trained, anticipation error is set as 0.0003, learning rate is 0.3, the additional momentum factor is 0.5, and frequency of training is set as 5000 times, trains the final connection weights of complete preservation network;

The in-vehicle sound quality evaluation model can be expressed as:

$\mathrm{SQ}=\underset{j=1}{\overset{14}{\mathrm{\&Sigma;}}}[{\mathrm{<figure-callout\; id="2"\; label="w\; j"\; filenames="HDA0000383635590000041.png"\; state="\{\{state\}\}">w</figure-callout>}}_j^{}\&times;\frac{1}{1+\exp -(\underset{i=1}{\overset{7}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}^1{x}_i+B_j^1)}]+b^2---\left(18\right)$

In formula, the sound quality subjective assessment value that SQ is internal car noise, i.e. irritated degree grade; x _i(i=1,2 ... 7) be the objective evaluation parameter of noise signal, be respectively shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index;

for input layer in table 2 to i parameter in the connection weights W1 of hidden layer to j neuronic connection value; for input layer j value of threshold value B1 to hidden layer; connect j the value of weights W2 to the sound quality value for hidden layer in table 2; b ²for the threshold value of hidden layer to the sound quality value.

Weight analysis described in technical scheme refers to the reconstruction evaluation model:

In concrete test, utilize the connection weights of neural network to calculate the method for input parameter to the output variable weighing factor, its mathematical computations equation is as follows:

$I_j=\frac{\underset{m=1}{\overset{\mathrm{Nh}}{\mathrm{\&Sigma;}}}(\left(\right|w_{\mathrm{jm}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right)}{\underset{k=1}{\overset{\mathrm{Ni}}{\mathrm{\&Sigma;}}}\{{\mathrm{\&Sigma;}}_{m=1}^{\mathrm{Nh}}\left(\right|w_{\mathrm{km}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right\}}---\left(19\right)$

In formula: I _jbe the weighing factor of j input parameter to n output variable; Ni, Nh is input and hidden layer node number; w ¹for the connection weights of input layer to hidden layer, w ²for the connection weights of hidden layer to output layer, corresponding subscript j, m, which neuron n is.As, be j input neuron and hidden layer m the neuronic weights that are connected.I _jbe worth greatlyr, the weighing factor ratio is higher, means that this input parameter is larger on the impact of output.

The weight coefficient of 7 objective psychoacoustic parameters of sound quality that calculate by formula (19) to the sound quality value, find out the high parameter of weight proportion in 7 objective psychoacoustic parameters, ignore influence coefficient little, then according to the method for the objective quantification model of setting up the in-vehicle sound quality subjective assessment, the objective quantification model of the subjective assessment of sound quality is optimized again.

Compared with prior art the invention has the beneficial effects as follows:

1. passenger car internal car noise sound quality prediction method of the present invention provides the method for a high efficient and reliable for the evaluation analysis of automobile sound quality.Simplified new car in the Acoustic Optimization design, repeatedly organize the professional repeatedly to carry out the loaded down with trivial details flow process that the subjective assessment of sound quality is analyzed, shortened for 90% test duration.The method only need to be carried out an infrasonic sound quality evaluation analysis and tests and set up evaluation model designing initial sample car, after completing again the further optimization improvement of experiment sample car, use this model to replace loaded down with trivial details time-consuming in-vehicle sound quality subjectivity to listen to test, only need the objective psychoacoustic parameter of in-vehicle sound quality of measuring and analyze appointment just can complete in-vehicle sound quality subjective assessment prediction, improved work efficiency, save manpower and materials, shortened the construction cycle of vehicle.

2. in passenger car internal car noise sound quality prediction method of the present invention, adopt dummy head or two microphones to simulate the perception of people's ear to noise signal with the method for binaural periphery computation model, both considered the emulation of whole method, making to predict the outcome more approaches people's subjective feeling, has improved the precision predicted the outcome of sound quality.Consider to use the economy of the method simultaneously, limited by experimental apparatus, under nobody foreman's prerequisite, also can substitute with microphone harmony quality binaural model, reach same experimental result, increase the versatility of the method, reduce it and use restriction.

3. the sound quality subjective assessment objective quantification model in passenger car internal car noise sound quality prediction method of the present invention adopts neural network (GA-BP) method of genetic algorithm optimization to set up, the global property good by genetic algorithm carrys out optimization neural network, make it there is self-evolution, adaptive ability and ability of searching optimum, and then make the more intelligent and robotization of whole Forecasting Methodology.

4. passenger car internal car noise sound quality prediction method of the present invention has reproduciblely, and the related coefficient of the result of its prediction and actual subjective assessment value reaches 0.95.

5. passenger car internal car noise sound quality prediction method highly versatile of the present invention, be widely used, and both can be applied to real train test, carries out the evaluation of sound quality prediction and and guide the sound Quality Design in car.Can also be basis with Virtual Design Platform, combine with the CAE technology, complete the Acoustic Optimization design process of whole automobile.

6. the objective psychoacoustic parameter database of sound quality in passenger car internal car noise sound quality prediction method of the present invention, be convenient in design to solve and the in-vehicle sound quality of other different brands same model vehicles mark is analyzed, further improve sound quality prediction appraisement system, improve the comparability predict the outcome, for the Acoustic Optimization validity of vehicle provides foundation.

The accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further illustrated:

Fig. 1 is new and old sound quality evaluation flow process contrast schematic block diagram of the present invention;

The overall procedure block diagram that Fig. 2 is passenger car internal car noise sound quality prediction method of the present invention;

The principle of work block diagram of the passenger car internal car noise sound quality prediction method that Fig. 3 is employing dummy head of the present invention;

The principle of work block diagram of the passenger car internal car noise sound quality prediction method that Fig. 4 is two microphones of employing of the present invention+binaural periphery computation model;

The structural principle block diagram of the sound quality binaural periphery computation model that Fig. 5 is passenger car internal car noise sound quality prediction method of the present invention;

The evaluation model that Fig. 6 is passenger car internal car noise sound quality prediction method of the present invention is set up the operating process block diagram;

The neural model structural drawing of Fig. 7 passenger car internal car noise of the present invention sound quality prediction method;

The evaluation model structural drawing of Fig. 8 passenger car internal car noise of the present invention sound quality prediction method;

The weight coefficient of 7 objective psychoacoustic parameters of Fig. 9 passenger car internal car noise of the present invention sound quality prediction method to sound quality evaluation value;

Figure 10 invents the FB(flow block) of the genetic algorithm of described passenger car internal car noise sound quality prediction method;

The structure principle chart of the binaural periphery computation model based on the Gammatone bank of filters of Figure 11 passenger car internal car noise of the present invention sound quality prediction method.

Embodiment

Below in conjunction with accompanying drawing, the present invention is explained in detail:

Referring to Fig. 1, the present invention, according in the past in the passenger car Acoustic Optimization, will carry out the objective psychoacoustic parameter analysis experiment of a set of loaded down with trivial details sound quality subjective assessment harmony quality for real vehicle transformation each time.A kind of passenger car in-vehicle sound quality subjective assessment Forecasting Methodology based on sound quality objective quantification model has been proposed.The present invention has simplified the experimental procedure that the passenger car in-vehicle sound quality is estimated, set up the relation of subjective assessment and objective noise testing data, the sound quality subjective assessment result after the Accurate Prediction sample car improves, optimize whole vehicle acoustic design process, shorten the construction cycle of vehicle, cost-saving.

One. the equipment and the computing method that in passenger car internal car noise sound quality prediction method of the present invention, adopt:

1. described dummy head is the manikin of the size of simulates real head part shoulder, and ear's structure of dummy head is the size of simulation true man ear, and sound transducer respectively is arranged in two ears.Can reduce truly auditor's sound field environment, reproduce external ear from physical form, collection, location and the amplification process of ear's physiological structures such as middle ear and inner ear to voice signal.

2. described microphone refers to sound transducer general on market, is commonly called as Mike.

3. described binaural periphery computation model is to adopt the method for mathematical computations to complete the sound signal processing process of ear to the identification cognitive process of sound:

1) transmitting procedure of sound from external ear to middle ear;

2) acoustical signal is decomposed into spectrum component;

3) post-processing stages: by the auditory physiology to auditory properties and inner ear, describe, sensitivity and the frequency selectivity of auditory system are enhanced.Can complete the physiology course through remarkable external ear, middle ear, inner ear and auditory nerve to ear sound by binaural periphery computation model.

Described IIR bandpass filter, Gammatone wave filter, low-pass filter, wide band filter group, half rectifier, nonlinear adaptive and in short-term integration be all that design of filter and analysis tool (FDATool) in the MATLAB Matlab DSPToolBox calculated design cell, can call easily change design parameter wherein.Wherein the IIR bandpass filter is iir digital filter, has another name called " infinite impulse response digital filter ", or " regressive filter ".There is feedback, it is generally acknowledged and there is unlimited impulse response.The Gammatone wave filter plays the time-domain filtering effect on bionics, single-channel voice is resolved into to the time-domain signal of a plurality of different frequency ranges, and this is equivalent to construct with single channel signal the multi channel signals effect of different frequency composition.

Two. passenger car internal car noise sound quality prediction method of the present invention

Consult Fig. 2, the step of passenger car internal car noise sound quality prediction method of the present invention is as follows:

1. utilize dummy head or two microphones to gather the internal car noise signal

Dummy head or two microphones are placed in to the copilot position, the internal car noise signal of collection vehicle under operating mode at the uniform velocity, and the internal car noise signal under operating mode is at the uniform velocity carried out to pre-service.

1), for the actual signal of objective comprehensive collection internal car noise, the optimum position that dummy head and two microphones are laid is the copilot position.The vertical coordinate that the copilot locates two positions that microphone is put is that seat surface center line and horizontal ordinate are the above 0.7 ± 0.05m of backrest surface center line intersection point places, two microphones are arranged in 0.2 ± 0.02m place, plane of symmetry left and right of seat surface and backrest symmetrically, and the position of dummy head's ear is identical with the position of microphone.On the one hand, for the car under operating mode at the uniform velocity, this position is nearest apart from Main Noise Sources (nacelle), made an uproar and all kinds of noise effect maximums such as road rumble by airborne noise, passing noise, wind.On the other hand, due to the experiment rules, position of driver needs person skilled to be operated vehicle, and the copilot position is nearest from the driver, is also to test the position of paying close attention to the most simultaneously.

2) tested object is selected the basis of the passenger car of different model different brands as the collecting test sample, quantity with the passenger car model is as the criterion, recommend passenger car brand and model in sample preferably to contain main flow passenger car vehicle on market, preferably guarantee the sample car of 3～4 kinds of brands of model selection of the same race, for example the A of car, B and C level car.

3) when test working condition requirement that noise signal is gathered, transmission for vehicles is put the most high-grade, take 10km/h～20km/h as a step-length, from the idling to the top speed, at the uniform velocity travels respectively, records the noise level of each speed of a motor vehicle in getting off.Each working condition requirement record 2 times, be greater than 30s writing time.

Sample sound to the internal car noise that gathers under each steady state condition is processed, and needs through segmentation, intercepting and screening, and the length legislations of each operating mode sample sound intercepting is the 5s left and right, and requires to be subject to the steady state noise signal of external interference minimum as far as possible.

More its last result of calculation of test sample book quantity is more accurate in theory, but consider the requirement of experimental cost and method computation complexity, suitable customization test sample book quantity is very important, recommended amount is in 70～150 left and right, at least guarantee 30, so not only the precision of ensuring method but also reduce workload.

2. calculate the objective psychoacoustic parameter of A sound level harmony quality

Consult Fig. 3 and Fig. 4, the objective psychoacoustic parameter of A sound level harmony quality that calculates internal car noise can adopt two kinds of methods to obtain, adopt dummy head or microphone and the binaural periphery computation model simulation people ear apperceive characteristic to sound, calculate the objective psychoacoustic parameter of A sound level harmony quality according to noise sample;

1) utilize Artemis sound attributional analysis software directly to calculate A sound level and objective psychoacoustic parameter (shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index).

2) adopt microphone and binaural periphery computation model to calculate the objective psychoacoustic parameter of A sound level harmony quality.By MATLAB the programme foundation that realizes binaural periphery computation model and the computation model of objective psychoacoustic parameter.Simulate auditory system to the sound signal processing process by computing method, join during the objective psychoacoustic parameter of sound quality calculates as the pretreatment stage of acoustical signal, can improve the performance of whole objective evaluation system.

Consult Fig. 5, the computational modeling of auditory periphery simulated sound is through the process of remarkable external ear, middle ear, inner ear and auditory nerve.The implementation procedure of binaural periphery computation model is divided two parts: being at first to set up monaural hearing periphery computation model (because the unfamiliar to the ear reason structure in left and right is identical), is then to calculate the calculating of ears loudness and relevant objective psychoacoustic parameter; Two each commissarial left and right ears of microphone, complete the collection to the internal car noise signal.It is as follows that it sets up monaural hearing periphery computation model concrete steps:

The first step: adopting microphone is for simulating the collection of whole external ear to voice signal, and voice signal is converted to digital signal.

Second step: because there is DC bias in the sample sound after gathering, cause having loudness differences between each acoustics sample.Adopt the iir filter group formed by 4 parallel IIR bandpass filter in MATLAB, can realize the special filter action of middle ear, use filtering direct current biasing component from input signal, eliminate the loudness differences that may exist between each acoustics sample, adjust the amplitude of acoustical signal, make it to reach the sound pressure level of appointment.

The 3rd step: because inner ear comprises corresponding nervous function, for the simulation of inner ear, at first adopt a broadband filter group to simulate the processing procedure of inner ear basement membrane to sound, recommend adoption Gammatone bank of filters or DRNL bank of filters realize.Because inner ear is for the acoustic processing rule difference of different frequency bands, when frequency is low, the amplitude peak of basement membrane vibration appears near the helicotrema place, and along with the raising of sound frequency, this peak value moves to the basement membrane root.So the voice signal frequency range of the iir filter group adopted output mainly is distributed in 20Hz～4kHz, the broadband filter group can be by the sound frequency division tape handling of different frequency.The low-pass first order filter that half-wave rectification and cutoff frequency are 1kHz is for simulating the processing procedure of auditory nerve to voice signal.Then the low-pass first order filter that is 1kHz by half-wave rectification, cutoff frequency simulates auditory nervous system, kept signal at the instantaneous microtexture of low frequency and extracted the envelope of signal at HFS.The signal of output just is transformed into the intensity characterization method that is similar to application square expansion.Last nonlinear adaptive, by the end of the low-pass filter of 8Hz and in short-term integration be for emulation the adaptive characteristic of sense of hearing periphery.Its objective is that voice signal does system-gain and change dynamically linearization more.

Consult Figure 11, what shown employing in figure is that the binaural that 64 Gammatone wave filters form is calculated the periphery model, centre frequency fixes on 20Hz～4kHz scope, corresponding Bark critical band number is 1～18, the wave filter spacing is 0.5Bark, and the time domain respective formula of its gammatone (GT) wave filter is:

In formula (1), n is the exponent number of wave filter, and B is bandwidth, f _ccentre frequency Hz,

be phase place, k is gain, and t is the time.

The 4th step: the voice signal that finally will process imports in the computing formula of the objective psychoacoustic parameter of each sound quality, calculates each parameter of the objective psychologic acoustics of each sound quality

The process of the objective psychoacoustic parameter of calculating in-vehicle sound quality is as follows:

(1) extract two microphones and gather the left and right otoacoustic emission frequency spectrum of arbitrarily downgrading, and it is carried out to frequency range analysis, calculate binaural sound by formula (2) and arbitrarily downgrade:

In formula, L _{p, L/R}(f) be binaural effect sound pressure level dB; The summation of A (p, f) left and right ear loudness; The absolute acoustic pressure P of p; F signal frequency Hz; L _p,Lthe left otoacoustic emission dB that arbitrarily downgrades; L _p,Rauris dextra sound pressure level dB.

Determine the vehicle interior sound field of each frequency band, and determine whether reverberation field.If reverberation field carries out the reverberation field attenuation processing, then carry out main loudness calculating, if not, just can direct main loudness calculate, slope loudness is calculated and finally can be derived ears characteristic loudness and total loudness, and it is as follows that it calculates the ears characteristic loudness:

$N^{\&prime;}\left(z\right)=0.08\left(\frac{E_{\mathrm{T\&Omega;}}}{E_0}\right)[{(0.5+0.5\frac{E}{E_{\mathrm{T\&Omega;}}})}^{0.23}-1]---\left(3\right)$

In formula, E _{t Ω}for excitation corresponding to the threshold of audibility under quiet situation; E ₀for reference sound intensity I ₀=10 ^-12w/m ²corresponding excitation; E is calculated the excitation that sound is corresponding.

Ears loudness computation model adopts the computation model of Zwicker, uses 1/3 octave band as basic data, and introduce the critical band concept masking effect of people's ear is done to corresponding being modified to,

$N=\underset{0}{\overset{24\mathrm{Bark}}{\&Integral;}}{N}^{\&prime;}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, be N'(z) characteristic loudness in critical band, unit is sone _g/ Bark, subscript G representation feature loudness value is calculated and is got by the critical band sound level.

By characteristic loudness, the sharpness of calculating sound quality, roughness and shake degree.

(2) the calculating sharpness is as follows:

$S=k\&times;\frac{{\&Integral;}_0^{24}{N}^{\&prime;}\left(z\right)g\left(z\right)\mathrm{zdz}}{N}\mathrm{acum}---\left(5\right)$

In formula, S represents sharpness, the acum of unit; K=0.11 is weighting coefficient; N is ears loudness; N ' is (z) characteristic loudness in critical band; Z is critical band Bark number; G (z) is the weighting function of Zwicker according to the different critical frequency band, and its analytic expression about the Bark territory of g (z) is

$g\left(z\right)=\left\{\begin{array}{ccc}1& \mathrm{for}& z\&le;16\\ {0.066e}^{0.171z}& \mathrm{for}& z>16\end{array}\right\}---\left(6\right)$

Z is critical band Bark number;

(3) the calculating roughness is as follows:

$R=0.3f_{\mathrm{mod}}{\&Integral;}_{z=0}^{24\mathrm{Bzrk}}{\mathrm{\&Delta;L}}_E\left(z\right)\mathrm{dzasper}---\left(7\right)$

In formula: R represents roughness, and unit is asper; f _modfor modulating frequency; Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage, be defined as:

${\mathrm{\&Delta;L}}_E\left(z\right)={20\log }_{10}\left(\frac{{N^{\&prime;}}_{\max }\left(z\right)}{{N^{\&prime;}}_{\min }\left(z\right)}\right)---\left(8\right)$

In formula: N ' _maxand N ' (z) _min(z) maximal value and the minimum value of difference representation feature loudness.

(4) calculating shake degree is as follows:

$F=0.008\frac{{\&Integral;}_0^{24}\mathrm{\&Delta;L}\left(z\right)\mathrm{dz}}{\left(\frac{f_{\mathrm{mod}}}{4}\right)+\left(\frac{4}{f_{\mathrm{mod}}}\right)}\mathrm{vacil}---\left(10\right)$

In formula: F representative shake degree, unit is vacil; f _modfor modulating frequency; Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage.

(5) calculating tone degree is as follows:

The analysis that the tone degree passes through each frequency band, in conjunction with the left and right feature band, obtains the related coefficient of adjacent feature frequency band, and obtains after calculating modulating frequency, and its computation model is as follows

$T=\sqrt{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left[W_1\left({\mathrm{\&Delta;z}}_i\right)W_2\left(f_1\right)W_3\left({\mathrm{\&Delta;L}}_i\right)\right]}^2}---\left(11\right)$

In formula, T represents the tone degree, and unit is tu; W ₁(Δ Z _i) mean the relation of i single-frequency composition and critical band difference; W ₂(f _i) mean the relation of i single-frequency composition and frequency; W ₃(Δ L _i) mean the sound level surplus graded effect of i single-frequency composition.

(6) calculating AI index is as follows:

According to the analysis to each frequency band, obtain the AI index, speech intelligibility (AI index) depends on sound pressure level and the frequency (1/3rd octaves) of ground unrest.The speech range that when people talk in car, people's ear can be listened shows as the zone of a 200～6300Hz in the spectrogram of ground unrest, and its computation model is

AI=∑W(f)D(f)/3

$\left\{\begin{array}{c}N\left(f\right)>\mathrm{UL}\left(f\right),D\left(f\right)=0\\ \mathrm{LL}\left(f\right)<N\left(f\right)<\mathrm{UL}\left(f\right),D\left(f\right)=\mathrm{UL}\left(f\right)-N\left(f\right)\\ N\left(f\right)<\mathrm{LL}\left(f\right),D\left(f\right)=30\end{array}\right.---\left(12\right)$

In formula, W (f) is the weighted coefficient; N (f) is noise spectrum; UL (f) is the ground unrest upper limit, and its expression formula is

lL (f) is the ground unrest lower limit, and its expression formula is the upper limit noise that its each frequency range is corresponding and weighted coefficient are as following table 1

The upper limit noise that each frequency range of table 1 is corresponding and weighted coefficient

Frequency/Hz	Upper limit noise UL (f)	Weighted coefficient W (f)
			200	64	1
250	69	2
			315	71	3.25
400	73	4.25

500	75	4.5
			630	75	5.25
800	75	6.5
			1000	74	7.25
1250	72	8.5
			1600	70	11.5
2000	67	11
			2500	65	9.5
3150	63	9
			4000	60	7.75
5000	56	6.25
			6300	51	2.5

3. in-vehicle sound quality subjective assessment

Noise sample is carried out to the subjective assessment of sound quality, obtain the subjective assessment result of test data.

1) in concrete test, consider automobile sound quality and there is multiple attribute, as Preference, irritated degree, luxurious sense, kinesthesia etc.There is very big difference in various crowds to the preference of vehicle sounds, and its influence factor has different zones, social class, age bracket, culture background and habits and customs etc.Subjective assessment can be chosen suitable sound quality evaluation index according to vehicle and crowd's needs, adopts the grade scoring method to carry out the subjective assessment test.Sound quality of the present invention subjective assessment is not limited to the method, also can adopt Paired Comparisons or grouped comparison method etc.

2), in the selection of estimating main body quantity, main body is estimated in 20 of minimum assurances, recommends 32～40 people.Requirement is selected from college student, the staff who is engaged in field of vibration noise, acoustics expert, automotive research personnel, testing crew, and Health Certificate is all without the hearing disease, and the age, the mean age was 30 years old between 22～50 years old, and M-F is 25:7; The evaluation personnel that choose have certain automobile noise experience, and accept to carry out certain training before the subjective assessment test.

3) on the subjective assessment sample process, for fear of the estimator due to experience with get sth into one's head and affect the judgement to sample sound, in the subjective assessment process of the test, should not make the estimator understand affiliated vehicle and the speed of a motor vehicle of each sample, before evaluation, each sample sound be sorted.

4. set up the objective quantification model of in-vehicle sound quality subjective assessment

Take A sound level in car, the objective psychoacoustic parameter of in-vehicle sound quality and sound quality subjective assessment result as basis, by the improved BP neural network of genetic algorithm, set up the objective quantification model of the subjective assessment of in-vehicle sound quality;

In concrete test, the sample sound that training gathers, each sample sound has comprised the objective psychoacoustic parameter of in-vehicle sound quality and subjective sound quality evaluation result after computational analysis, and sample set has been contained the car homeostasis noise under the different speed of a motor vehicle.The subjective irritated degree of sound quality of internal car noise of usining is exported as network, 7 objective psychoacoustic parameters of in-vehicle sound quality (shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index) are inputted as network, adopt genetic algorithm to improve the BP neural network and realize that the objective psychological sound parameter of in-vehicle sound quality arrives the mapping between subjective sound quality evaluation result, sets up the objective quantification model of in-vehicle sound quality subjective assessment.

Consult Fig. 6, the step of the objective quantification model of the in-vehicle sound quality subjective assessment of foundation based on the improved neural network of genetic algorithm (BP) is as follows:

1) determine initial configuration and the parameter of neural network (BP);

Consult Fig. 7, in concrete test, recommend 3 layers of neural network to build the objective quantification model of in-vehicle sound quality subjective assessment, input layer is the input vector X that 7 objective parameters (shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index) form, and nodes is 7; Output layer is subjective sound quality evaluation score value, and nodes is 1.The hidden layer node number is mainly by the purposes decision of network, and suitable number does not have clear and definite theory provision, and the experimental formula of reference is as follows:

$m=\sqrt{n+l}+\mathrm{\&alpha;}---\left(13\right)$

m=log ₂n （14）

$\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{C}_m^i>k---\left(15\right)$

In above-mentioned 3 formulas: m is the hidden layer neuron number; N is the input layer number; L is the output layer neuron number; A, i is constant; K is sample number.Hidden layer adopts the logsig function, and output layer adopts the purlin linear function, and the neural network of structure (BP) model structure as shown in Figure 7.

2) determine neural network (BP) operation, parameters, encoded

Introducing genetic algorithm (GA) operative algorithm is optimized neural network (BP).Select the encoding scheme of genetic algebra, variation probability, population scale and employing.The length of chromosome coding is determined by each layer of contained neuron number.If the input layer of neural network (BP) has R neuron, hidden layer has S1 neuron, and output layer neuron number is S2, code length S=R*S1+S1*S2+S1+S2.

3) carry out genetic algorithm (GA) operation according to fitness;

Fitness function determines the learning error of population training by sample, and learning error is expressed as follows:

$E\left(i\right)=\underset{n}{\mathrm{\&Sigma;}}\underset{m}{\mathrm{\&Sigma;}}{(y_m-o_m)}^2---\left(16\right)$

In above formula, i is the chromosome number, and m is the output node number, and n is number of training.The fitness function of genetic algorithm can be expressed as:

f(i)=1/E(i) （17）

In above formula, f (i) is each individual fitness function.

According to the flow process of genetic algorithm (GA), as Figure 10, utilize the GA tool box of Matlab to be solved, and the individual optimal solution obtained is decoded, as the initial network weights of neural network.During decoding, the corresponding relation of each real number and BP neural network weight, threshold value is as shown in table 2.

Table 2 decoding relation

W1	W2	B1	B2
				R*S1	S1*S2	S1	S2

In table 2, W1 be neural network (BP) input layer to the connection weights between hidden layer, W2 be hidden layer to the connection weights between output layer, B1 is the hidden layer threshold value, B2 is the output layer threshold value.

4) using the optimization weights that obtain as the initial network weights, then use genetic algorithm (GA) neural network training (BP);

After adopting genetic algorithm (GA) optimization acquisition initial network weights and threshold value, just can carry out neural network (BP) training by setup parameter.The setting of concrete training parameter is most important to neural network (BP), the performance of its network (BP) that directly affect the nerves.Adopt the additional momentum method to be trained, anticipation error is set as 0.0003, and learning rate is 0.3, and the additional momentum factor is 0.5, and frequency of training is set as 5000 times.Train the final connection weights of complete preservation neural network (BP).

Consult Fig. 7, connection weights and the threshold value of neural network (BP) have represented the connection performance that it is inner, i.e. the parameter of constructed mathematical model, and the in-vehicle sound quality evaluation model can be expressed as:

$\mathrm{SQ}=\underset{j=1}{\overset{14}{\mathrm{\&Sigma;}}}[{\mathrm{<figure-callout\; id="2"\; label="w\; j"\; filenames="HDA0000383635590000041.png"\; state="\{\{state\}\}">w</figure-callout>}}_j^{}\&times;\frac{1}{1+\exp -(\underset{i=1}{\overset{7}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}^1{x}_i+B_j^1)}]+b^2---\left(18\right)$

In formula, the sound quality subjective assessment value that SQ is internal car noise, i.e. irritated degree grade; x _i(i=1,2 ... 7) be the objective evaluation parameter 1 of noise signal, be respectively shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index; for input layer in table 2 to i parameter in the connection weights W1 of hidden layer to j neuronic connection value;

for input layer j value of threshold value B1 to hidden layer;

connect j the value of weights W2 to the sound quality value for hidden layer in table 2; b ²for the threshold value of hidden layer to the sound quality value;

Weight analysis with rebuild evaluation model

By weight analysis, find out the objective parameter that the subjective assessment result is had to material impact, take the objective parameter of material impact and sound quality subjective assessment result is basis, rebuilds the objective quantification model of the subjective assessment of the sound quality under steady state condition.

In concrete test, utilize the connection weights of neural network to calculate the method for input parameter to the output variable weighing factor, its mathematical computations equation is as follows:

$I_j=\frac{\underset{m=1}{\overset{\mathrm{Nh}}{\mathrm{\&Sigma;}}}(\left(\right|w_{\mathrm{jm}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right)}{\underset{k=1}{\overset{\mathrm{Ni}}{\mathrm{\&Sigma;}}}\{{\mathrm{\&Sigma;}}_{m=1}^{\mathrm{Nh}}\left(\right|w_{\mathrm{km}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right\}}---\left(19\right)$

In formula: I _jbe the weighing factor of j input parameter to n output variable; Ni, Nh is input and hidden layer node number; w ¹for the connection weights of input layer to hidden layer, w ²for the connection weights of hidden layer to output layer, corresponding subscript j, m, which neuron n is.As,

be j input neuron and hidden layer m the neuronic weights that are connected.I _jbe worth greatlyr, the weighing factor ratio is higher, means that this input parameter is larger on the impact of output.

The weight coefficient of 7 objective psychoacoustic parameters of sound quality that calculate by formula (19) to the sound quality value, find out the high parameter of weight proportion in 7 objective psychoacoustic parameters, ignore influence coefficient little, then according to the method for step 4, the objective quantification model of the subjective assessment of sound quality is optimized again.

6. set up the objective psychoacoustic data of in-vehicle sound quality storehouse;

Set up the objective psychoacoustic data of passenger car in-vehicle sound quality storehouse, set up the relation of objective quantification model of the subjective assessment of passenger car in-vehicle sound quality objective psychoacoustic data storehouse and sound quality, form fast and convenient sound quality prediction method.

In concrete test, set up the database of the objective psychoacoustic parameter of sound quality of different automobile types, as the data basis of whole passenger car internal car noise sound quality prediction method.Its database is convenient to solve in-vehicle sound quality prediction mark is analyzed in the vehicle acoustic design is optimized, and further improves sound quality prediction appraisement system, for the method is set up the basic platform of a system.

Embodiment

Understand this method for ease of person skilled is more deep, describe by the concrete case study on implementation that the method is applied to domestic car sound quality prediction:

1. in order to obtain abundant data sample, choosing the A with representative, B and C level domestic car fast-selling on market is tested object, and every kind of vehicle is selected the same model car of 4 domestic representative brands.The test site that noise signal gathers is at the smooth asphalt road of car load semianechoic room and suburb.During outdoor test, select vehicular traffic few, carry out when external interference is few.

The dummy head is placed in to the copilot position, and transmission for vehicles is put the most high-grade, with idling, 40km/h, 60km/h, 80km/h, 100km/h, 120km/h, at the uniform velocity travels respectively, records the noise level of each speed of a motor vehicle in getting off.Each driving cycle record 2 times, be greater than 30sec writing time.To the get off sample sound of interior noise of each steady state condition, through segmentation, intercepting, screening, finally obtained 78 sample sounds, the length of each operating mode sample intercepting is the 5s left and right, and to each sample degree of making a sound equalization

2. utilize Artemis sound attributional analysis software directly to calculate A sound level and the objective psychoacoustic parameter (shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index) of each noise sample

3. selected 32 sound evaluation personnel altogether from college student, the staff who is engaged in field of vibration noise, acoustics expert, automotive research personnel, testing crew, Health Certificate is all without the hearing disease.Estimator's age, the mean age was 30 years old between 22～50 years old; Wherein, M-F is 25:7; The evaluation personnel that choose have certain automobile noise experience, and accept to carry out certain training before the subjective assessment test.

Focus on the characteristics of the comfortable stationarity of vehicle according to Chinese, select irritated degree as sound quality evaluation index.Select the grade scoring method to carry out subjective sound quality evaluation test.The agitation degree is divided into 11 grades by international standard, as shown in table 4.

The table 4 subjective assessment agitation degree grade table of comparisons

Very

Bad

Very poor

Poor

Dissatisfied

Can accept

Satisfied

Better

Good

Fine

Fabulous

1

2

3

4

5

6

7

8

9

10

11

Before carrying out evaluation test, the testing crew that each is participated in estimating by operating personnel illustrates detailed step and the points for attention of evaluation method, and some problems that may occur in process of the test are explained.In process of the test, each estimator can hear different noise sample successively, according to its subjective sensation, presses respectively corresponding digital button on scoring device.Such as, after the estimator listens the noise sample of finishing the trail, if he thinks that this noise sample makes us sensation " better ", just can press the button " 8 " and give a mark accordingly scoring device is enterprising; If need to again give a mark, click the C key and can return.The estimator clicks "enter" key" after result is determined, just completes the marking of a sample sound.In the subjective test process, after certain noise sample is play, if the estimator can not make clear and definite marking to sample sound, can again to this sample, be listened to evaluation, until obtain satisfied result.

Evaluation result to all evaluation personnel is carried out data detection, calculates the average correlation coefficient between each estimator.Use SPSS software to calculate each evaluation personnel Spearman coefficient each other; Then every estimator and other all estimators' related coefficient is got to arithmetic equal value, calculate average correlation coefficient.Average correlation coefficient between each estimator is as shown in table 5, rejects the relatively low personnel's scoring of related coefficient.

Each estimator's of table 5 average correlation coefficient

The estimator	1	2	3	4	5	6	7	8	9	10	11
												Related coefficient	0.786	0.819	0.56	0.83	0.82	0.8	0.61	0.781	0.75	0.862	0.804
The estimator	12	13	14	15	16	17	18	19	20	21	22
												Related coefficient	0.438	0.783	0.843	0.58	0.865	0.779	0.78	0.845	0.816	0.54	0.782
The estimator	23	24	25	26	23	27	28	29	30	31	32
												Related coefficient	0.85	0.779	0.752	0.823	0.85	0.487	0.76	0.84	0.791	0.81	0.85

Remaining evaluation result is carried out to the K-mean cluster analysis.Cluster analysis result is as shown in table 6, the first kind 4 examples, and Equations of The Second Kind 20 examples, lack 2 examples.Therefore estimator's number of the first kind very little, can not represent most of estimators' result, selects 20 estimators' of Equations of The Second Kind appraisal result to calculate the subjective irritated degree value of sound quality of each sample sound.

Table 6 cluster analysis result

4. the objective quantification model of in-vehicle sound quality subjective assessment is set up

Directly calculate the objective psychoacoustic parameter of in-vehicle sound quality by the Artemis sound attributional analysis software in step 2, the subjective assessment of each test sample book of table 7 expression is the objective psychoacoustic parameter of harmony quality as a result.

The subjective and objective result of table 7 sample

Consult Fig. 8, select 3 layers of neural network to build the evaluation model of sound quality, input layer is the input vector X that 7 objective parameters (shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index) form, and nodes is 7; Output layer is subjective sound quality evaluation score value, and nodes is 1.Selected hidden layer node number is 14.Thereby the network topology structure built is 7-14-1, hidden layer adopts the logsig function, and output layer adopts the purlin linear function.

Consult Fig. 7, before the formal training neural network, introduce the GA algorithm BP network is optimized, genetic algebra is chosen as 150, and the variation probability is 0.05, and population scale is 50.Coded system is the real number method, the BP network structure that the present invention sets up is 7-14-1, according to table 2 and described hereditary code length S=R*S1+S1*S2+S1+S2, the neuron number that R is input layer, S1 is the hidden layer neuron number, and S2 is output layer neuron number.So code length S=127.According to the flow process of genetic algorithm, as Figure 10, utilize the GA tool box of Matlab to be solved, and the individual optimal solution obtained is decoded, as the initial network weights of neural network.During decoding, the corresponding relation of each real number and BP neural network weight, threshold value is as shown in table 2.Genetic algorithm training is complete, and it is as shown in table 8 to weights and the threshold value of hidden layer to carry out decoded input layer, and hidden layer is as shown in table 9 to weights and the threshold value of output layer, and network model as shown in Figure 7.

Table 8 input layer is to weights and the threshold value of hidden layer

Table 9 hidden layer is to weights and the threshold value of output layer

After obtaining the initial weight and threshold value of neural network, adopt the additional momentum method to be trained, anticipation error is set as 0.0003, and learning rate is 0.3, and the additional momentum factor is 0.5, and frequency of training is set as 5000 times.Training sample is the subjective and objective data in table 4.

To sum up the complete subjective result of each sample training and objective psychoacoustic parameter and weights and Network of Threshold are as shown in table 10.And by the objective quantification model of mathematical model (18) foundation in-vehicle sound quality subjective assessment as shown in Figure 8.

Table 10 weights and threshold matrix

5. weight analysis and evaluation model optimization

Consult Fig. 9, bring the interior data of table 10 into formula (19), the weight coefficient of 7 objective psychoacoustic parameters that calculate to the sound quality value.Therefrom can find out: the parameter that sound quality subjective assessment result is had the greatest impact is ears loudness, is secondly roughness and sharpness.These three parameters have reached 83% to the contribution degree sum of sound quality impact, and other several objective parameter impacts are less.

Three parameters of loudness, roughness, sharpness that then will material impact be arranged to the sound quality are as input, and the subjective result of sound quality, for output, rebuilds sound quality GA-BP in-vehicle sound quality subjective assessment objective quantification model.The structure of neural network (BP) is defined as 3-7-1, and learning parameter, genetic algorithm (GA) Optimization Steps, training and check data remain unchanged, and complete Model Reconstruction.Follow-up internal car noise is carried out to the sound attributional analysis while estimating, the sound quality evaluation model that application is set up, only need three psychology objective parameter of calculating noise, substitution sound quality evaluation model is calculated, with regard to the measurable subjectivity agitation degree value that obtains noise sample, thereby saved loaded down with trivial details time-consuming subjectivity and listened to test, simplified flow process.

6. set up the database of the objective psychoacoustic parameter of different automobile types in-vehicle sound quality, as whole passenger car internal car noise sound quality prediction method data basis.Its database is convenient to solve in-vehicle sound quality prediction mark is analyzed in the vehicle acoustic design is optimized, and further improves sound quality prediction appraisement system, for the method is set up the basic platform of a system.

Claims (5)

1. a passenger car internal car noise sound quality prediction method, is characterized in that, the step of described passenger car internal car noise sound quality prediction method is as follows:

1) utilize dummy head or two microphones to gather the internal car noise signal:

Dummy head or two microphones are placed in to the copilot position, the internal car noise signal of collection vehicle under operating mode at the uniform velocity, and the internal car noise signal under operating mode is at the uniform velocity carried out to pre-service;

2) calculate the objective psychoacoustic parameter of A sound level harmony quality:

Adopt computational modeling of auditory periphery to calculate the objective psychoacoustic parameter of in-vehicle sound quality;

3) in-vehicle sound quality subjective assessment:

Noise sample is carried out to the subjective assessment of sound quality, obtain the subjective assessment result of test data;

4) set up the objective quantification model of in-vehicle sound quality subjective assessment:

Take A sound level in car, the objective psychoacoustic parameter of in-vehicle sound quality and sound quality subjective assessment result as basis, by the improved BP neural network of genetic algorithm, set up the objective quantification model of the subjective assessment of in-vehicle sound quality;

5) weight analysis and reconstruction evaluation model:

Find out the objective parameter that the subjective assessment result is had to material impact by weight analysis, take the objective parameter of material impact and sound quality subjective assessment result is basis, rebuilds the objective quantification model of the subjective assessment of the sound quality under steady state condition;

6) set up the objective psychoacoustic data of in-vehicle sound quality storehouse:

In concrete test, set up the database of the objective psychoacoustic parameter of sound quality of different automobile types, as the data basis of whole passenger car internal car noise sound quality prediction method; Database is convenient to solve in-vehicle sound quality prediction mark is analyzed in the vehicle acoustic design is optimized, and further improves sound quality prediction appraisement system, for passenger car internal car noise sound quality prediction method is set up the basic platform of a system.

2. according to passenger car internal car noise sound quality prediction method claimed in claim 1, it is characterized in that, the described step of dummy head or two microphone collection internal car noise signals of utilizing is as follows:

1) dummy head or two microphones are placed on to the copilot place, the vertical coordinate that is dummy head or two microphone particular locations is that seat surface center line and horizontal ordinate are the above 0.7 ± 0.05m of backrest surface center line intersection point places, and dummy head or two microphones are arranged in 0.2 ± 0.02m place about the plane of symmetry of seat surface and backrest symmetrically;

2) tested object is selected the basis of the passenger car of different model different brands as the collecting test sample, quantity with the passenger car model is as the criterion, passenger car brand and model contain main flow passenger car vehicle on market, guarantee the sample car of 3～4 kinds of brands of model selection of the same race, as A, B and the C level car of car;

The working condition requirement while 3) testing, noise signal gathered, transmission for vehicles is put the most high-grade, take 10km/h～20km/h as a step-length, from the idling to the top speed, at the uniform velocity travel respectively, record the noise level of each speed of a motor vehicle in getting off, each working condition requirement record 2 times, be greater than 30s writing time;

Sample sound to the internal car noise that gathers under each steady state condition is processed, and needs through segmentation, intercepting and screening, and the length legislations of each operating mode sample sound intercepting is the 5s left and right, and requires to be subject to the steady state noise signal of external interference minimum as far as possible;

More its last result of calculation of test sample book quantity is more accurate, considers the requirement of testing cost and method computation complexity, and quantity, in 70～150 left and right, at least guarantees 30.

3. according to passenger car internal car noise sound quality prediction method claimed in claim 1, it is characterized in that, the step of the objective psychoacoustic parameter of described employing computational modeling of auditory periphery calculating in-vehicle sound quality is as follows:

(1) extract two microphones and gather the left and right otoacoustic emission frequency spectrum of arbitrarily downgrading, and it is carried out to frequency range analysis, calculate binaural sound by formula (2) and arbitrarily downgrade:

In formula, L _{p, L/R}(f) be binaural effect sound pressure level dB; The summation of A (p, f) left and right ear loudness; The absolute acoustic pressure P of p; F signal frequency Hz; L _p,Lthe left otoacoustic emission dB that arbitrarily downgrades; L _p,Rauris dextra sound pressure level dB;

Determine the vehicle interior sound field of each frequency band, and determine whether reverberation field, if reverberation field, carry out the reverberation field attenuation processing, carry out again main loudness calculating, if not, with regard to direct main loudness, calculate, slope loudness is calculated and is finally derived ears characteristic loudness and total loudness, and its ears characteristic loudness is as follows:

$N^{\&prime;}\left(z\right)=0.08\left(\frac{E_{\mathrm{T\&Omega;}}}{E_0}\right)[{(0.5+0.5\frac{E}{E_{\mathrm{T\&Omega;}}})}^{0.23}-1]---\left(3\right)$

In formula, E _{t Ω}for excitation corresponding to the threshold of audibility under quiet situation; E ₀for reference sound intensity; I ₀=10 ^-12w/m ²corresponding excitation; E is calculated the excitation that sound is corresponding;

Ears loudness computation model adopts the computation model of Zwicker, uses 1/3 octave band as basic data, and introduce the critical band concept masking effect of people's ear is done to corresponding being modified to,

$N=\underset{0}{\overset{24\mathrm{Bark}}{\&Integral;}}{N}^{\&prime;}\left(z\right)\mathrm{dz}---\left(4\right)$

In formula, be N'(z) characteristic loudness in critical band, unit is sone _g/ Bark, subscript G representation feature loudness value is calculated and is got by the critical band sound level;

(2) the calculating sharpness is as follows:

$S=k\&times;\frac{{\&Integral;}_0^{24}{N}^{\&prime;}\left(z\right)g\left(z\right)\mathrm{zdz}}{N}\mathrm{acum}---\left(5\right)$

In formula, S represents sharpness, the acum of unit; K=0.11 is weighting coefficient; N is ears loudness; N ' is (z) characteristic loudness in critical band; Z is critical band Bark number; G (z) is the weighting function of Zwicker according to the different critical frequency band, and its analytic expression about the Bark territory of g (z) is

$g\left(z\right)=\left\{\begin{array}{ccc}1& \mathrm{for}& z\&le;16\\ {0.066e}^{0.171z}& \mathrm{for}& z>16\end{array}\right\}---\left(6\right)$

Z is critical band Bark number;

(3) the calculating roughness is as follows:

$R=0.3f_{\mathrm{mod}}{\&Integral;}_{z=0}^{24\mathrm{Bzrk}}{\mathrm{\&Delta;L}}_E\left(z\right)\mathrm{dzasper}---\left(7\right)$

In formula: R represents roughness, and unit is asper; f _modfor modulating frequency; Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage, be defined as:

${\mathrm{\&Delta;L}}_E\left(z\right)={20\log }_{10}\left(\frac{{N^{\&prime;}}_{\max }\left(z\right)}{{N^{\&prime;}}_{\min }\left(z\right)}\right)---\left(8\right)$

In formula: N ' _maxand N ' (z) _min(z) maximal value and the minimum value of difference representation feature loudness;

(4) calculating shake degree is as follows:

$F=0.008\frac{{\&Integral;}_0^{24}\mathrm{\&Delta;L}\left(z\right)\mathrm{dz}}{\left(\frac{f_{\mathrm{mod}}}{4}\right)+\left(\frac{4}{f_{\mathrm{mod}}}\right)}\mathrm{vacil}---\left(10\right)$

In formula: F representative shake degree, unit is vacil; f _modfor modulating frequency, Δ L _e(z) be the variable quantity of acoustical signal drive(r) stage;

(5) analysis that the tone degree passes through each frequency band, in conjunction with the left and right feature band, obtains the related coefficient of adjacent feature frequency band, and obtains after calculating modulating frequency, and its computation model is as follows

$T=\sqrt{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left[W_1\left({\mathrm{\&Delta;z}}_i\right)W_2\left(f_1\right)W_3\left({\mathrm{\&Delta;L}}_i\right)\right]}^2}---\left(11\right)$

In formula, T represents the tone degree, and unit is tu; W ₁(Δ Z _i) mean the relation of i single-frequency composition and critical band difference; W ₂(f _i) mean the relation of i single-frequency composition and frequency; W ₃(Δ L _i) mean the sound level surplus graded effect of i single-frequency composition;

(6) obtain the AI index according to the analysis to each frequency band, speech intelligibility is sound pressure level and the frequency (1/3rd octaves) that the AI index depends on ground unrest, the speech range that when people talk in car, people's ear can be listened shows as the zone of a 200～6300Hz in the spectrogram of ground unrest, and its computation model is

AI=∑W(f)D(f)/30

$\left\{\begin{array}{c}N\left(f\right)>\mathrm{UL}\left(f\right),D\left(f\right)=0\\ \mathrm{LL}\left(f\right)<N\left(f\right)<\mathrm{UL}\left(f\right),D\left(f\right)=\mathrm{UL}\left(f\right)-N\left(f\right)\\ N\left(f\right)<\mathrm{LL}\left(f\right),D\left(f\right)=30\end{array}\right.---\left(12\right)$

In formula, W (f) is the weighted coefficient; N (f) is noise spectrum; UL (f) is the ground unrest upper limit, and its expression formula is

lL (f) is the ground unrest lower limit, and its expression formula is

4. according to passenger car internal car noise sound quality prediction method claimed in claim 1, it is characterized in that, the step of the described objective quantification model of setting up the in-vehicle sound quality subjective assessment is as follows:

1) determine initial configuration and the parameter of BP NEURAL NETWORK;

In concrete test, recommend 3 layers of neural network to build the objective quantification model of in-vehicle sound quality subjective assessment, input layer is the input vector X that shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and 7 objective parameters of AI index form, and nodes is 7; Output layer is subjective sound quality evaluation score value, and nodes is 1; The hidden layer node number is mainly by the purposes decision of network, and suitable number does not have clear and definite theory provision, and the experimental formula of reference is as follows:

$m=\sqrt{n+l}+\mathrm{\&alpha;}---\left(13\right)$

m=log ₂n （14）

$\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{C}_m^i>k---\left(15\right)$

In above-mentioned 3 formulas: m is the hidden layer neuron number; N is the input layer number; L is the output layer neuron number; A, i are constant; K is sample number; Hidden layer adopts the logsig function, and output layer adopts the purlin linear function to build neural model;

2) determine BP NEURAL NETWORK operation, parameters, encoded

Introduce the Genetic Algorithms operative algorithm BP NEURAL NETWORK is optimized, select the encoding scheme of genetic algebra, variation probability, population scale and employing; The length of chromosome coding is by each layer of contained neuron number decision, and the input layer of establishing BP NEURAL NETWORK has R neuron, and hidden layer has S1 neuron, and output layer neuron number is S2, code length S=R*S1+S1*S2+S1+S2;

3) carry out the Genetic Algorithms operation according to fitness

Fitness function determines the learning error of population training by sample, and learning error is expressed as follows:

$E\left(i\right)=\underset{n}{\mathrm{\&Sigma;}}\underset{m}{\mathrm{\&Sigma;}}{(y_m-o_m)}^2---\left(16\right)$

In above formula, i is the chromosome number, and m is the output node number, and n is number of training.The fitness function of Genetic Algorithms can be expressed as:

f(i)=1/E(i) （17）

In above formula, f (i) is each individual fitness function;

Utilize the GA tool box of Matlab to be solved, and the individual optimal solution obtained is decoded, as the initial network weights of neural network, the corresponding relation of each real number and BP NEURAL NETWORK weights, threshold value during decoding;

4) using the optimization weights that obtain as the initial network weights, then use genetic algorithm neural network training BP;

After adopting Genetic Algorithms optimization acquisition initial weight and threshold value, just can carry out neural metwork training by setup parameter, adopt the additional momentum method to be trained, anticipation error is set as 0.0003, learning rate is 0.3, the additional momentum factor is 0.5, and frequency of training is set as 5000 times, trains the final connection weights of complete preservation network;

The in-vehicle sound quality evaluation model can be expressed as:

$\mathrm{SQ}=\underset{j=1}{\overset{14}{\mathrm{\&Sigma;}}}[w_j^2\&times;\frac{1}{1+\exp -(\underset{i=1}{\overset{7}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}^1{x}_i+B_j^1)}]+b^2---\left(18\right)$

In formula, the sound quality subjective assessment value that SQ is internal car noise, i.e. irritated degree grade; x _i(i=1,2 ... 7) be the objective evaluation parameter of noise signal, be respectively shake degree, sharpness, tone degree, A sound level, ears loudness, roughness and AI index;

for input layer in table 2 to i parameter in the connection weights W1 of hidden layer to j neuronic connection value;

for input layer j value of threshold value B1 to hidden layer; connect j the value of weights W2 to the sound quality value for hidden layer in table 2; b ²for the threshold value of hidden layer to the sound quality value.

5. according to passenger car internal car noise sound quality prediction method claimed in claim 1, it is characterized in that, described weight analysis refers to the reconstruction evaluation model:

In concrete test, utilize the connection weights of BP NEURAL NETWORK to calculate the method for input parameter to the output variable weighing factor, its mathematical computations equation is as follows:

$I_j=\frac{\underset{m=1}{\overset{\mathrm{Nh}}{\mathrm{\&Sigma;}}}(\left(\right|w_{\mathrm{jm}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right)}{\underset{k=1}{\overset{\mathrm{Ni}}{\mathrm{\&Sigma;}}}\{{\mathrm{\&Sigma;}}_{m=1}^{\mathrm{Nh}}\left(\right|w_{\mathrm{km}}^1|/{\mathrm{\&Sigma;}}_{k=1}^{\mathrm{Ni}}|w_{\mathrm{km}}^1\left|\right)\&times;|w_{\mathrm{mn}}^2\left|\right\}}---\left(19\right)$

In formula: I _jbe the weighing factor of j input parameter to n output variable; Ni, Nh is input and hidden layer node number; w ¹for the connection weights of input layer to hidden layer, w ²for the connection weights of hidden layer to output layer, corresponding subscript j, m, which neuron n is.As,

be j input neuron and hidden layer m the neuronic weights that are connected.I _jbe worth greatlyr, the weighing factor ratio is higher, means that this input parameter is larger on the impact of output.

The weight coefficient of 7 objective psychoacoustic parameters of sound quality that calculate by formula (19) to the sound quality value, find out the high parameter of weight proportion in 7 objective psychoacoustic parameters, ignore influence coefficient little, then according to the method for the objective quantification model of setting up the in-vehicle sound quality subjective assessment, the objective quantification model of the subjective assessment of sound quality is optimized again.

Images